This paper proposes a smart elderly care behavior anomaly detection method based on multimodal sensor data fusion and lightweight deep convolutional networks to address issues such as high misclassification rates in complex home environments and difficulties in aligning cross-modal features inherent in traditional single-modal perception. The framework first synchronizes 30 fps visual data with 100 Hz inertial and environmental sensor streams through a spatio-temporal alignment module. It then employs an enhanced EfficientNet-B0 backbone, utilizing 2D-CNN to extract visual spatial features and 1D-CNN to capture inertial and environmental temporal characteristics. A cross-modal attention mechanism dynamically weights contributions from the three modalities, enabling adaptive fusion of heterogeneous features. Finally, weighted cross-entropy loss mitigates the scarcity of abnormal samples, enabling real-time binary classification and four-category anomaly segmentation at the edge. Experiments on the self-built CARE-Home dataset (50 subjects, 360 hours, 4,400 anomalies) and the public UR Fall dataset demonstrate an F1 score of 88.4% and an AUC of 0.924, representing a 4.9% improvement over the best baseline LSTM fusion. For fall detection, the F1 score reaches 94.0%. Ablation studies confirm that spatio-temporal alignment and attention modules contribute 3.1% and 1.5% F1 gains respectively. With only 4.9 million parameters, the model runs at 30 FPS with 15 W power consumption on NVIDIA Jetson AGX Xavier, meeting the low-power, high-accuracy, real-time anomaly monitoring requirements for home-based elderly care.
Jiang et al. (Sun,) studied this question.